1-1
1 INTRODUCTION
Microarray technology simultaneously measures the mRNA of tens of thousands of genes in biological samples in a high-throughput and cost- effective manner. Since its introduction in 1995 [1], microarray technology has improved dramatically and became a widely used tool to study the whole transcriptome of many organisms. It has been adopted to explore the molecular basis of fundamental biological processes and complex diseases [2, 3], to improve the disease taxonomy [4, 5], to classify patients into known disease subclasses [6], to analyze the response to drug administration [7], and to predict disease outcomes [8, 9].
Enhancements in microarray technology and its widespread use have led to the generation of a relevant amount of data and resulted in several large public data repositories such as Gene Expression Omnibus (GEO) [10]
(http://www.ncbi.nlm.nih.gov/geo/) from NCBI, ArrayExpress [11]
(http://www.ebi.ac.uk/arrayexpress/) from EBI and CIBEX (Center for Information Biology gene EXpression database) [12] (http://cibex.nig.ac.jp/).
1-2
It is not uncommon to find multiple microarray gene expression studies performed by different research groups worldwide addressing the same or similar biological questions. Hence there has been a growing interest in developing methods to efficiently integrate microarray data from independent studies with the aim of fully exploiting the rich information produced. Meta- analysis appears to be an effective solution to this pressing issue [13].
As stated by Hedges, “meta-analysis consists of statistical methods for combining results from independent but related studies” [14]. However the term meta-analysis is also widely used in a broader sense, as we do here, to indicate the whole process of identification, selection, assessment and quantitative synthesis of several studies concerning a well-defined research question [15]. Many people use the term meta-analysis interchangeably with systematic review, however not all the systematic reviews are meta-analyses.
In fact a meta-analysis is a systematic review which provides a statistical synthesis of the results and produces an overall estimate of the effect of interest.
Meta-analysis offers several practical advantages.
First of all, meta-analysis represents an inexpensive solution to overcome the problem of reduced statistical power of microarray experiments and to reveal true effects of interest [16]. Typically, in microarray experiments many probes are investigated in few samples due to the high cost of this technology or the lack of biological replicates available. The straight consequence is that studies with small sample sizes are less likely to detect true effects and more prone to false positive and false negative results. Putting results together, therefore, increases the sample size and the statistical power of the study. It also allows a more accurate estimation of the effect, even if derived from small but consistent variations.
1-3
Moreover, meta-analysis has the potential to strengthen and extend the results obtained by individual studies and to increase their reliability. Indeed, it has been shown that microarray studies are poorly reproducible across platforms and/or laboratories [17, 18]. Technological differences among different microarray platforms [19], large variations in biological and experimental settings, small sample sizes and inappropriate statistical methods [20, 21] have been pointed out as the major sources that contribute to the inconsistency of microarray results. Many of these can be assessed and controlled or overcome by the use of standard reporting methods and the careful application of large-scale meta-analysis techniques with an appropriate statistical modeling of the inter-study variation [16].
Meta-analysis has been widely used in the area of medical and epidemiological research as well as in the sociological and behavioral sciences [22]. The applicability of meta-analysis methods to microarray datasets was demonstrated for the first time in 2002 by Rhodes who combined four datasets on prostate cancer to determine genes that were differentially expressed between clinically localized prostate tumor and benign prostate tissue samples [23]. Since then, several applications of meta-analysis to microarray data appeared in the literature [24-26].
Through a systematic search on PubMed, Tseng and colleagues [27] found that 333 microarray meta-analysis papers (including reviews, biological applications, methodological articles and database/software description papers) were published until December 2010, thus confirming the relevant interest of the scientific community in this challenging task. In more than half of the above mentioned publications, meta-analysis was applied to identify Differentially Expressed Genes (DEGs) between two or more conditions [28-30].
However microarray studies have also been combined for classification analysis
1-4
[31], to identify co-expressed genes or to build gene networks [32-34], to evaluate reproducibility and bias across studies [35-37]. Figure 1.1 illustrates a microarray meta-analyses summary performed by Tseng and colleagues [27].
Figure 1.1: Classification of the 333 microarray meta-analysis papers reviewed by Tseng based on the type of paper (A) and the purpose of meta-analysis (B) (image modified from [27])
Despite its increasing popularity, however, meta-analysis of microarray data is not without problems. In fact, although it shares many features with traditional meta-analysis, most classical meta-analysis methods cannot be directly applied to microarray experiments because of their unique issues such as the large number of variables involved and the technical complexities of combining data across different experimental platforms (e.g. gene nomenclatures, species and analytical methods) [38].
1.1 AIM OF THE STUDY
In this study, we focused on the application of meta-analysis to the two- class comparison microarray experiments. The objective of this kind of studies is to identify DEGs between two well-defined conditions, namely cases and controls. Four statistical approaches were comparatively evaluated: the weighted version of the inverse normal method by Marot and Mayer [39] and the moderated effect size combination approach proposed by Marot [40], both
1-5
implemented in the R (http://www.r-project.org/) package metaMA, the t- based hierarchical modeling described in Choi et al. [16] and implemented in the Bioconductor (http://www.bioconductor.org/) package GeneMeta [41] and the rank product method with the RankProd Bioconductor package [42]. These methods were applied to a set of three publicly available microarray studies on malignant pleural mesothelioma to identify DEGs between normal and malignant mesothelioma pleural tissues. Since it is not yet clear if filtering is beneficial from a meta-analysis perspective, both unfiltered and filtered data were analyzed to evaluate the impact of a common filtering strategy on meta- analysis results.
